stock-analyzer-cli

An interactive equity-research CLI. It runs an AI agent (on the Pi coding-agent framework) that produces structured, citation-backed reports from live SEC EDGAR and Yahoo Finance data — every figure is fetched through a typed tool before the model writes, so nothing in a report is recalled from memory.

Each report is a drop-in Markdown "skill" invoked as a slash command: /moat NVDA, /decide AAPL.

Disclaimer

This is a research tool, not investment advice. Reports are analytical framing, not buy/sell recommendations, and no output should be relied on for a financial decision.

Data sources. Filings and financials come from SEC EDGAR, the authoritative source. Prices and analyst consensus come from Yahoo Finance endpoints that are unofficial and undocumented — Yahoo retired its public finance API in 2017. These endpoints are not licensed for this use, are accessed contrary to Yahoo's Terms of Service, and may change or break without notice. Use this for personal research only; do not build a commercial product or a hosted service on the Yahoo-backed tools.

Consensus is opinion. Analyst estimates, price targets, ratings and the earnings calendar are third-party opinion, not filed fact, and are labelled as such in every report. /story and /moat read only SEC data.

Provided "as is", without warranty of any kind — see LICENSE.

Requirements

• Node.js 22.19+, required by the Pi agent framework. On older Node, npm silently resolves Pi to its legacy-node20 tag instead of failing, so check node --version first.

• An API key for the agent model — Anthropic, OpenAI, Google, xAI, DeepSeek or OpenRouter. No key is needed for the data: EDGAR and Yahoo are used without auth.

• A contact address for SEC EDGAR, which its fair-access policy requires on every request. The CLI asks once on first run and saves it. As a Pi package there is no prompt, so you must set it yourself:

  export SEC_USER_AGENT="Your Name you@example.com"
  

  There is no placeholder default — without an address, EDGAR requests refuse to send rather than go out unidentified, so every report fails. See docs/data-sources.md.

Install

Two ways to run this, with the same reports and the same data tools either way.

As a standalone CLI

npm install -g stock-analyzer-cli
stock-analyze                 # interactive session
stock-analyze NVDA moat       # straight into a report

Inside Pi

If you already use pi, this is the shortest path — pi supplies the model, key management and session loop, so there is nothing to configure:

pi install npm:stock-analyzer-cli
pi -e npm:stock-analyzer-cli     # or try it once, without installing

Every report is then a slash command in your pi session, aliases included.

From source

git clone git@github.com:vsambaraju/stock-analyzer-cli.git
cd stock-analyzer-cli
npm install                   # also builds, via the prepare script
npm link                      # optional: puts `stock-analyze` on your PATH

dist/ is git-ignored build output; npm run build rebuilds it after editing sources.

API keys

Credentials resolve in this order:

1. Environment variables (table below), or Pi's ~/.pi/agent/auth.json
2. ~/.stock-analyzer/config.json (owner-only, 0600)
3. An interactive first-run wizard, only if no provider is configured

Keys are never written into the project directory and never committed.

Provider	--provider	Env var
Anthropic	anthropic	ANTHROPIC_API_KEY
OpenAI	openai	OPENAI_API_KEY
Google Gemini	google	GEMINI_API_KEY
xAI (Grok)	xai	XAI_API_KEY
DeepSeek	deepseek	DEEPSEEK_API_KEY
OpenRouter	openrouter	OPENROUTER_API_KEY
Azure OpenAI Responses	azure-openai-responses	AZURE_OPENAI_API_KEY (+ AZURE_OPENAI_BASE_URL)
Amazon Bedrock	amazon-bedrock	AWS_BEARER_TOKEN_BEDROCK or AWS profile/IAM

With no flags the CLI auto-selects the first configured provider (anthropic → openai → google → xai → deepseek → openrouter → azure → bedrock) and a sensible default model:

stock-analyze --provider google TSLA      # use Gemini
stock-analyze --model grok-4.5 NVDA moat  # a specific model
stock-analyze --list-models               # configured providers + their models

Inside a session, /model lists and switches models live; /model <query> filters (/model gemini) or matches an id directly (/model gpt-4o).

Usage

Command	Action
/<report> [TICKER]	Run a report on the current stock, or on TICKER if given
/moat [TICKER] [PEER…]	Durability report, peer-checked against companies you name
/new [TICKER]	Switch to a different stock
/model [query]	List/switch the AI model across configured providers
/help	List all reports and controls
/exit	Quit
(anything else)	A free-form follow-up, answered with the tools + prior reports

A dim footer after each turn reports token usage and tool-call count for that prompt, and the session total on /new or exit. It is omitted if the model runtime doesn't report usage.

Reports

Five reports, ordered as a research pass. Every report name you may have used before survives as an alias pointing at whichever report now owns it.

Command	Aliases	What it does
/story	business, biz, longterm, lt, growth, tam	What the business is, where growth comes from, and whether its markets are expanding
/moat	durability, risk, saas, ai, compete, peers	Moat size and direction as a pair, the four threats to it, and an optional peer check
/numbers	phase, metrics, scorecard	Lifecycle phase (1–5) plus the Red/Yellow/Green scorecard for that phase
/earnings	quarter, results, print, catalysts, events	The latest print, what management guided, and the near-term calendar
/decide	qst, valuation, value, sentiment, price, timing	Quality · Size · Timing — the verdict, carrying the full valuation and the stage read

Run them in that order — what it is, whether it lasts, whether it's executing, what just happened, then the verdict. /decide is the only one that answers whether to own it; /earnings is the one built for quarterly re-runs on a name you already know.

📖 docs/reports.md — what each report does and why: the shared moat rubric, the three questions, the four-stage technical read, the lifecycle classifier, and how peers and TAM are handled.

How it works

• src/tools/ — the typed data tools the agent calls:
  • market.ts — Yahoo price data/history, the weekly four-stage get_technical_stage read, SEC XBRL financials, and the get_business_phase classifier
  • filings.ts / edgar.ts — EDGAR filing lookup, section extraction (10-K/10-Q Item 1, 1A, 7, 7A), the decoded 8-K event log, identity and rate limiting
  • estimates.ts — Yahoo consensus, earnings calendar, 90-day estimate revisions
  • segments.ts — per-segment revenue and operating income, plus peer comparison
  • guidance.ts — guidance and quarterly segment highlights from the 8-K item 2.02 earnings release
• src/skills/*.md — the report protocols. Frontmatter (name, order, aliases, description, optional kickoffHint and args) plus a prompt body. Drop a new .md file here and it becomes a /command — no code changes needed.
• src/skills-shared/*.md — partials pulled into more than one report by a {{include: moat-rubric.md}} line. moat-rubric.md holds the moat and threat criteria that both /moat and /decide score against — one copy, included verbatim, so the two reports cannot rate the same filings differently. Edit the rubric here and nowhere else. These live outside src/skills/ because Pi turns every .md under that directory into a command of its own, and a partial is not a command.
• src/extension.ts — registers the tools with the Pi agent.
• src/cli.ts — the interactive session, argument parsing, and system prompt.

🔌 docs/data-sources.md — where each figure comes from and what these sources cannot do: the SEC identity requirement, why segment data can't come from companyfacts, how guidance is read out of an 8-K exhibit, why trailing multiples differ from a quote site, and what is deliberately unavailable.

Development

npm run dev      # run from source with tsx
npm run build    # tsc, then expand skill includes and chmod the CLI
npm start        # run the built CLI (node dist/cli.js)

Both entry points share everything that matters: src/extension.ts registers the data tools, src/skills/*.md holds the report protocols, and buildReportMessage() in src/commands.ts composes the prompt for both, so a report reads the same either way.

src/pi-aliases.ts is the one file that is not shared. It registers each skill's aliases: as slash commands, because Pi ignores unknown frontmatter fields. It is referenced only by the pi manifest in package.json — never from cli.ts, which resolves aliases through its own lookup table. Loading it in both places would register two competing handlers for every alias.

Skill frontmatter must be valid YAML. Pi parses it with a real YAML parser and silently skips any skill whose frontmatter fails to parse, so values with special characters need quoting — e.g. args: "[TICKER] [PEER ...]".

License

MIT © Venkat Sambaraju